This dissertation focuses on the development of an intelligent framework based on Agentic Large Language Models (Agentic LLMs), applied to the analysis, comprehension, and, ultimately, the automatic restoration of complex historical documents. The overall objective is to propose an architecture capable of processing highly heterogeneous documents by relying on specialized and collaborative agents, capable of breaking down complex tasks and implementing progressive, structured, and explainable reasoning.
In the first research area, the focus is on in-depth understanding of historical documents, particularly on modeling their logical, visual, and semantic structures. These documents have specific characteristics such as irregular layouts, marginal annotations, figures, tables, and implicit spatial relationships between the various elements. The goal is to identify, segment, and organize these components into a structured and usable representation, using multimodal approaches that combine computer vision and natural language processing.

A second focus of the dissertation concerns the restoration and reconstruction of degraded content. Since historical documents are often affected by physical damage (faded ink, tears, incomplete pages, or missing fragments), the goal is to develop methods capable of reconstructing missing information while ensuring linguistic, stylistic, and historical consistency. This section relies in particular on mechanisms of conditional generation, contextual completion, and reasoning guided by the documentary context.
The thesis also includes the design of a multi-agent architecture based on Agentic LLMs, in which different specialized agents are responsible for complementary tasks such as visual analysis, semantic interpretation, multimodal fusion, and content generation. The interaction and coordination among these agents aim to improve the system’s robustness, accuracy, and generalization ability in the face of the diversity and complexity of historical documents.
An important cross-cutting aspect of the work concerns the integration of explainability mechanisms (Explainable AI). The goal is to ensure the transparency and interpretability of the proposed framework by providing traceability of the decisions made by the agents, justification of the results produced, and visualization of the intermediate steps in the processing workflow.

Finally, the thesis calls for the implementation of a rigorous evaluation protocol that combines quantitative metrics (precision, recall, F1-score, correct reconstruction rate) with qualitative assessments conducted in collaboration with experts in the field (historians, linguists, archivists). This dual approach makes it possible to evaluate both the technical performance and the historical and semantic relevance of the results obtained.